Efficient 3D Reconstruction and Real-Time Rendering of Large Environments Using Gaussian Splatting

Sammanfattning

Gaussian Splatting is a technique that enables photorealistic 3D reconstructions directly from photographic data, but applying the technique to large-scale indoor environments on consumer-grade hardware introduces significant computational and memory challenges. No robust pipeline has previously been demonstrated that handles data acquisition, reconstruction and real-time rendering of a complete indoor environment at this scale using only consumer-grade hardware. This thesis, conducted in collaboration with Stadium, presents, evaluates and develops a batch-based reconstruction pipeline applied to their distribution center DC100 in Norrköping, with a total area of 58,000 square meters. An analysis of time and memory complexity identified the number of simultaneously processed images as the dominant limiting factor, motivating a division of DC100 into 66 independent sections. A dedicated image acquisition strategy was developed for large-scale indoor environments and the individually reconstructed sections were merged into a coherent model via shared camera parameters. To enable real-time rendering, post-processing optimizations were applied to reduce computational demands and a distance-based level streaming system was implemented in Unreal Engine 5. Following completed reconstruction, based on 23,722 images with a total size of 79.5 GB, the work resulted in a digital model of 3 GB that despite low FPS could be navigated in real time on a laptop with integrated graphics, while on a dedicated consumer graphics card it performed at over 200 FPS. The results demonstrate that the developed method enables large-scale reconstructions with Gaussian Splatting that can be generated and rendered in real time using only consumer-grade hardware, for both data acquisition, reconstruction and real-time rendering. Future work for the method includes improved robustness in environments with repetitive structures, integration of external position data to strengthen reconstruction quality, and further optimization to enable smooth real-time rendering on low-end consumer hardware.